With the increasing popularity of the Internet, social media plays a crucial role in people's daily communication. However, due to the anonymity of Internet, toxic comments emerge in an endless stream on the Internet, which seriously affects the health of online social environment. To effectively reduce the impact of toxic comments, automated scoring methods for the severity of toxic comments are in great demand. For that purpose, a deep-learning-based natural language processing technique is proposed using ELECTRA to automatically score the toxicity of a comment in this work. The backbone of our model is the ELECTRA discriminator, and the downstream regression task is accomplished by the following head layer. Three head layers are implemented separately: multi-layer perceptron, convolutional neural network, and attention. The dataset used for model training is from the Kaggle competition Toxic Comment Classification Challenge, and the model performance is evaluated through another Kaggle competition Jigsaw Rate Severity of Toxic Comments. By a boost from the K-Fold cross validation and an ensemble of three models with different head layers, our method can reach a competition score 0.80343. Such score ranks 71/2301 (top 3.1%) in the leaderboard and can get a silver medal in the competition. The results in this work would help filter the toxic comments and harmful text information automatically and effectively on the Internet, and could greatly reduce the cost of manual review and help build a healthier Internet environment.
{"title":"Deep-Learning-Based Automated Scoring for the Severity of Toxic Comments Using Electra","authors":"Tiancong Zhang","doi":"10.1145/3556677.3556693","DOIUrl":"https://doi.org/10.1145/3556677.3556693","url":null,"abstract":"With the increasing popularity of the Internet, social media plays a crucial role in people's daily communication. However, due to the anonymity of Internet, toxic comments emerge in an endless stream on the Internet, which seriously affects the health of online social environment. To effectively reduce the impact of toxic comments, automated scoring methods for the severity of toxic comments are in great demand. For that purpose, a deep-learning-based natural language processing technique is proposed using ELECTRA to automatically score the toxicity of a comment in this work. The backbone of our model is the ELECTRA discriminator, and the downstream regression task is accomplished by the following head layer. Three head layers are implemented separately: multi-layer perceptron, convolutional neural network, and attention. The dataset used for model training is from the Kaggle competition Toxic Comment Classification Challenge, and the model performance is evaluated through another Kaggle competition Jigsaw Rate Severity of Toxic Comments. By a boost from the K-Fold cross validation and an ensemble of three models with different head layers, our method can reach a competition score 0.80343. Such score ranks 71/2301 (top 3.1%) in the leaderboard and can get a silver medal in the competition. The results in this work would help filter the toxic comments and harmful text information automatically and effectively on the Internet, and could greatly reduce the cost of manual review and help build a healthier Internet environment.","PeriodicalId":350340,"journal":{"name":"Proceedings of the 2022 6th International Conference on Deep Learning Technologies","volume":"25 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126173202","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
After Covid-19 swept the globe and bitcoin prices suddenly soared, machine learnings were used to predict the trend of bitcoin prices, but these studies were lack of performance analysis in different time-scale span. In this paper, three neural network models are designed and used to forecast the price of bitcoin after the outbreak of COVID-19. The models A uses the high/low price, open/close price of four-days of bitcoin as input variables and the close price of the fifth day as target variable, the models B uses same variable as the model A and uses optimal weights, and the model C uses same structure as the model B, but adds the trading volume to the input variables. The results show that the model C may lower the difference between actual and calculated outputs, thus boosting the prediction accuracy. Also, it is found that the models that can work well when predicting bitcoin prices in a short time span can be obviously less precise when it comes to predicting bitcoin prices in a longer time span.
{"title":"Prediction of Bitcoin Price Since COVID-19 by Using Neural Network Models","authors":"Zhiheng Jiang","doi":"10.1145/3556677.3556679","DOIUrl":"https://doi.org/10.1145/3556677.3556679","url":null,"abstract":"After Covid-19 swept the globe and bitcoin prices suddenly soared, machine learnings were used to predict the trend of bitcoin prices, but these studies were lack of performance analysis in different time-scale span. In this paper, three neural network models are designed and used to forecast the price of bitcoin after the outbreak of COVID-19. The models A uses the high/low price, open/close price of four-days of bitcoin as input variables and the close price of the fifth day as target variable, the models B uses same variable as the model A and uses optimal weights, and the model C uses same structure as the model B, but adds the trading volume to the input variables. The results show that the model C may lower the difference between actual and calculated outputs, thus boosting the prediction accuracy. Also, it is found that the models that can work well when predicting bitcoin prices in a short time span can be obviously less precise when it comes to predicting bitcoin prices in a longer time span.","PeriodicalId":350340,"journal":{"name":"Proceedings of the 2022 6th International Conference on Deep Learning Technologies","volume":"21 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131762691","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In the early stage of fire, smoke alarm detection is an important means to prevent fire. And with the continuous construction of monitoring facilities, it is of great significance for the study of smoke video monitoring. In order to meet the detection accuracy and speed of the video, the EfficientDet target detection algorithm was improved. Firstly, the visual analysis of the smoke data set was carried out by clustering method, and the anchor frame parameters in the EfficientDet algorithm were re-calibrated by K-means clustering method. Secondly, the Bi-FPN feature fusion algorithm is improved to reduce the transverse cross-layer connection and increase the longitudinal cross-layer connection, which reduces the calculation of parameters and improves the detection accuracy. Finally, in order to solve the problem of missing detection in small smoke area, a two-channel attention mechanism is added to improve the detection accuracy.
{"title":"Smoke Detection Algorithm Based on Improved EfficientDet","authors":"Zengquan Yang, Han Huang, Fuming Xia, Zhen Shi","doi":"10.1145/3556677.3556678","DOIUrl":"https://doi.org/10.1145/3556677.3556678","url":null,"abstract":"In the early stage of fire, smoke alarm detection is an important means to prevent fire. And with the continuous construction of monitoring facilities, it is of great significance for the study of smoke video monitoring. In order to meet the detection accuracy and speed of the video, the EfficientDet target detection algorithm was improved. Firstly, the visual analysis of the smoke data set was carried out by clustering method, and the anchor frame parameters in the EfficientDet algorithm were re-calibrated by K-means clustering method. Secondly, the Bi-FPN feature fusion algorithm is improved to reduce the transverse cross-layer connection and increase the longitudinal cross-layer connection, which reduces the calculation of parameters and improves the detection accuracy. Finally, in order to solve the problem of missing detection in small smoke area, a two-channel attention mechanism is added to improve the detection accuracy.","PeriodicalId":350340,"journal":{"name":"Proceedings of the 2022 6th International Conference on Deep Learning Technologies","volume":"36 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131001289","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Vrushali Atul Surve, Pramod Pathak, Mohammed Hasanuzzaman, Rejwanul Haque, Paul Stynes
Classification of e-commerce products involves identifying the products and placing those products into the correct category. For example, men’s Nike Air Max will be in the men’s category shoes on an e-Commerce platform. Identifying the correct classification of a product from hundreds of categories is time-consuming for businesses. This research proposes an Image-based Transfer Learning Framework to classify the images into the correct category in the shortest time. The framework combines Image-based algorithms with Transfer Learning. This research compares the time to predict the category and accuracy of traditional CNN and transfer learning models such as VGG19, InceptionV3, ResNet50, and MobileNet. A visual classifier is trained CNN and transfer learning models such as VGG19, InceptionV3, ResNet50, and MobileNet. The models are trained on an e-commerce product dataset that combines the ImageNet dataset with pre-trained weights. The dataset consists of 15000 images scraped from the web. Results demonstrate that Inception V3 outperforms all other models based on a TIMING of 0.10 seconds and an accuracy of 85%.
{"title":"An Image-based Transfer Learning Framework for Classification of E-Commerce Products","authors":"Vrushali Atul Surve, Pramod Pathak, Mohammed Hasanuzzaman, Rejwanul Haque, Paul Stynes","doi":"10.1145/3556677.3556689","DOIUrl":"https://doi.org/10.1145/3556677.3556689","url":null,"abstract":"Classification of e-commerce products involves identifying the products and placing those products into the correct category. For example, men’s Nike Air Max will be in the men’s category shoes on an e-Commerce platform. Identifying the correct classification of a product from hundreds of categories is time-consuming for businesses. This research proposes an Image-based Transfer Learning Framework to classify the images into the correct category in the shortest time. The framework combines Image-based algorithms with Transfer Learning. This research compares the time to predict the category and accuracy of traditional CNN and transfer learning models such as VGG19, InceptionV3, ResNet50, and MobileNet. A visual classifier is trained CNN and transfer learning models such as VGG19, InceptionV3, ResNet50, and MobileNet. The models are trained on an e-commerce product dataset that combines the ImageNet dataset with pre-trained weights. The dataset consists of 15000 images scraped from the web. Results demonstrate that Inception V3 outperforms all other models based on a TIMING of 0.10 seconds and an accuracy of 85%.","PeriodicalId":350340,"journal":{"name":"Proceedings of the 2022 6th International Conference on Deep Learning Technologies","volume":"173 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123170021","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In low-voltage residential electricity scenarios, simple identification algorithms are difficult to be effective because of the many types of appliances and similar power characteristics. We propose a household load identification method based on multi-label and convolutional neural networks (ML-CNN) to address these problems. Firstly, we analyze the V-I trajectory characteristics of different loads and use the binary images of V-I trajectory mapping as the study features. Secondly, we collect the original steady-state voltage and current data of the combined operation of common household appliances and build a dataset. Finally, we pre-process and multi-label the dataset and input it into the ML-CNN network structure for training and validation. The experimental results show that the average identification accuracy of the ML-CNN method is 97.63%, which is better than the load identification methods such as multi-label k-nearest neighbor (ML-KNN) and support vector machine (SVM).
{"title":"Household Load Identification Based on Multi-label and Convolutional Neural Networks","authors":"Zhengquan Wang, Qi Xie","doi":"10.1145/3556677.3556695","DOIUrl":"https://doi.org/10.1145/3556677.3556695","url":null,"abstract":"In low-voltage residential electricity scenarios, simple identification algorithms are difficult to be effective because of the many types of appliances and similar power characteristics. We propose a household load identification method based on multi-label and convolutional neural networks (ML-CNN) to address these problems. Firstly, we analyze the V-I trajectory characteristics of different loads and use the binary images of V-I trajectory mapping as the study features. Secondly, we collect the original steady-state voltage and current data of the combined operation of common household appliances and build a dataset. Finally, we pre-process and multi-label the dataset and input it into the ML-CNN network structure for training and validation. The experimental results show that the average identification accuracy of the ML-CNN method is 97.63%, which is better than the load identification methods such as multi-label k-nearest neighbor (ML-KNN) and support vector machine (SVM).","PeriodicalId":350340,"journal":{"name":"Proceedings of the 2022 6th International Conference on Deep Learning Technologies","volume":"53 9","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"120971271","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Nowadays, omics datasets have been widely used to study cancer and other related problems, but there are many cancer subtypes in some types of cancer, and some types have not been studied, so we must use unsupervised methods for cluster analysis. Cluster analysis is the process of finding similar data points in a pile of data points and classifying them. In this paper, five omics data sets are used to compare the three clustering methods, in order to find a more suitable clustering method for omics datasets. The conclusion of this paper is that OPTICS method is a better clustering method.
{"title":"Comparison of cancer classification algorithms based on clustering analysis","authors":"Jiawei Guo, Yu-shan Cai","doi":"10.1145/3556677.3556684","DOIUrl":"https://doi.org/10.1145/3556677.3556684","url":null,"abstract":"Nowadays, omics datasets have been widely used to study cancer and other related problems, but there are many cancer subtypes in some types of cancer, and some types have not been studied, so we must use unsupervised methods for cluster analysis. Cluster analysis is the process of finding similar data points in a pile of data points and classifying them. In this paper, five omics data sets are used to compare the three clustering methods, in order to find a more suitable clustering method for omics datasets. The conclusion of this paper is that OPTICS method is a better clustering method.","PeriodicalId":350340,"journal":{"name":"Proceedings of the 2022 6th International Conference on Deep Learning Technologies","volume":"28 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132615908","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Ocean mesoscale eddy detection is an important hotspot of Marine scientific research. Over the last few years, with the development of machine learning research, eddy detection methods based on machine learning have been applied in various fields. However, the traditional scroll detection algorithm has weak generalization ability and low detection accuracy, and the fully supervised scroll detection algorithm needs a large amount of marker data, which is costly and has poor readability. In this paper, a new semi-supervised ocean mesoscale eddy detection method based on feature invariance is proposed. The fully supervised loss calculation model is optimized to solve the problem of serious imbalance of positive and negative samples in loss calculation, so as to achieve the purpose of training the model. In addition, based on the feature invariance, an interpolation consistency calculation method based on flipped image and original image is proposed, which is combined with the consistency method algorithm put forward in CSD networks to increase the precision of detection. Compared with SSD and ISD networks, the proposed meso-scale eddy detection algorithm achieves better performance, with the AP value increasing by 1.7% and 1.1%, respectively.
{"title":"Semi supervised ocean mesoscale vortex detection method based on feature invariance","authors":"Haiyan Liu, Bo Qin, Y. Liu","doi":"10.1145/3556677.3556682","DOIUrl":"https://doi.org/10.1145/3556677.3556682","url":null,"abstract":"Ocean mesoscale eddy detection is an important hotspot of Marine scientific research. Over the last few years, with the development of machine learning research, eddy detection methods based on machine learning have been applied in various fields. However, the traditional scroll detection algorithm has weak generalization ability and low detection accuracy, and the fully supervised scroll detection algorithm needs a large amount of marker data, which is costly and has poor readability. In this paper, a new semi-supervised ocean mesoscale eddy detection method based on feature invariance is proposed. The fully supervised loss calculation model is optimized to solve the problem of serious imbalance of positive and negative samples in loss calculation, so as to achieve the purpose of training the model. In addition, based on the feature invariance, an interpolation consistency calculation method based on flipped image and original image is proposed, which is combined with the consistency method algorithm put forward in CSD networks to increase the precision of detection. Compared with SSD and ISD networks, the proposed meso-scale eddy detection algorithm achieves better performance, with the AP value increasing by 1.7% and 1.1%, respectively.","PeriodicalId":350340,"journal":{"name":"Proceedings of the 2022 6th International Conference on Deep Learning Technologies","volume":"55 4 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123659024","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Rohan Indrajeet Jadhav, Paul Stynes, Pramod Pathak, Rejwanul Haque, Mohammed Hasanuzzaman
Categorizing of clothes from wild fashion images involves identifying the type of clothes a person wears from non-studio images such as a shirt, trousers, and so on. Identifying the fashion clothes from wild images that are often grainy, unfocused, with people in different poses is a challenge. This research proposes a comparison between object detection and instance segmentation based models to categorise clothes from wild fashion images. The Object detection model is implemented using Faster Region-Based Convolutional Neural Network (RCNN). Mask RCNN is used to implement an instance segmentation model. We have trained the models on standard benchmark dataset namely deepfashion2. Results demonstrate that Instance Segmentation models such as Mask RCNN outperforms Object Detection models by 20%. Mask RCNN achieved 21.05% average precision, 73% recall across the different IoU (Intersection over Union). These results show promise for using Instance Segmentation models for faster image retrieval based e-commerce applications.
从疯狂的时尚图片中对衣服进行分类包括从非工作室图片(如衬衫、裤子等)中识别一个人穿的衣服类型。从杂乱无章的照片中识别时尚服装是一项挑战，这些照片往往是颗粒状的，没有聚焦，人们摆出不同的姿势。本文提出了一种基于对象检测和实例分割的服装分类模型的比较方法。目标检测模型采用基于更快区域的卷积神经网络(RCNN)实现。掩码RCNN是用来实现实例分割模型的。我们在标准基准数据集deepfashion2上训练了模型。结果表明，Mask RCNN等实例分割模型的性能比目标检测模型高出20%。掩模RCNN在不同IoU (Intersection over Union)上的平均准确率达到21.05%，召回率达到73%。这些结果显示了使用实例分割模型来实现更快的基于图像检索的电子商务应用的前景。
{"title":"An Instance Segmentation Model to Categorize Clothes from Wild Fashion Images","authors":"Rohan Indrajeet Jadhav, Paul Stynes, Pramod Pathak, Rejwanul Haque, Mohammed Hasanuzzaman","doi":"10.1145/3556677.3556690","DOIUrl":"https://doi.org/10.1145/3556677.3556690","url":null,"abstract":"Categorizing of clothes from wild fashion images involves identifying the type of clothes a person wears from non-studio images such as a shirt, trousers, and so on. Identifying the fashion clothes from wild images that are often grainy, unfocused, with people in different poses is a challenge. This research proposes a comparison between object detection and instance segmentation based models to categorise clothes from wild fashion images. The Object detection model is implemented using Faster Region-Based Convolutional Neural Network (RCNN). Mask RCNN is used to implement an instance segmentation model. We have trained the models on standard benchmark dataset namely deepfashion2. Results demonstrate that Instance Segmentation models such as Mask RCNN outperforms Object Detection models by 20%. Mask RCNN achieved 21.05% average precision, 73% recall across the different IoU (Intersection over Union). These results show promise for using Instance Segmentation models for faster image retrieval based e-commerce applications.","PeriodicalId":350340,"journal":{"name":"Proceedings of the 2022 6th International Conference on Deep Learning Technologies","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128832441","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The increasing Distributed Denial of Service (DDoS) attacks on the Internet of Things (IoT) is leading to the need for an efficient detection approach. Although much research has been conducted to detect DDoS attacks on traditional networks, such as machine learning (ML) based approaches that have improved accuracy and confidence, the limited bandwidth and computation resources in IoT networks restrict the application of ML, especially deep learning (DL) based solutions that require extensive input data. In order to appropriately address the security issues in the resources-constrained IoT network, this paper is aimed to reduce the input data dimensions by extracting a subset of the most relevant features from the original features and using this subset to detect DDoS attacks on IoT without degrading the detection performance. A cost-effective model is developed to clean and prepare raw data before dimensionality reduction. A hybrid feature selection that uses Mutual Information (MI), Analysis of Variance (ANOVA), Chi-Squared, L1-based feature selection, and Tree-based feature selection algorithms is designed to identify important data features and reduce the data inputs needed for detection. Simulation results show that detection accuracy is improved with the combination of features chosen by the proposed hybrid feature selection approach. The training time is much less than the combination of each individual feature selection method.
{"title":"Hybrid Feature Selection for Efficient Detection of DDoS Attacks in IoT","authors":"Liang Hong, Khadijeh Wehbi, Tulha Hasan Alsalah","doi":"10.1145/3556677.3556687","DOIUrl":"https://doi.org/10.1145/3556677.3556687","url":null,"abstract":"The increasing Distributed Denial of Service (DDoS) attacks on the Internet of Things (IoT) is leading to the need for an efficient detection approach. Although much research has been conducted to detect DDoS attacks on traditional networks, such as machine learning (ML) based approaches that have improved accuracy and confidence, the limited bandwidth and computation resources in IoT networks restrict the application of ML, especially deep learning (DL) based solutions that require extensive input data. In order to appropriately address the security issues in the resources-constrained IoT network, this paper is aimed to reduce the input data dimensions by extracting a subset of the most relevant features from the original features and using this subset to detect DDoS attacks on IoT without degrading the detection performance. A cost-effective model is developed to clean and prepare raw data before dimensionality reduction. A hybrid feature selection that uses Mutual Information (MI), Analysis of Variance (ANOVA), Chi-Squared, L1-based feature selection, and Tree-based feature selection algorithms is designed to identify important data features and reduce the data inputs needed for detection. Simulation results show that detection accuracy is improved with the combination of features chosen by the proposed hybrid feature selection approach. The training time is much less than the combination of each individual feature selection method.","PeriodicalId":350340,"journal":{"name":"Proceedings of the 2022 6th International Conference on Deep Learning Technologies","volume":"290 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121264942","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The tracking analysis of ocean feature phenomena exists many problems, such as incomplete topological structure information extraction and unclear time-varying law information display, etc. In this paper, a topology-oriented 3D ocean flow field feature classification and tracking algorithm is proposed to solve the problem of flow field feature tracking in different scales. The algorithm consists of three parts: Initially, the adaptive circular sampling space manner is optimized and improved to adapt to the extraction of flow field feature regions at different scales in view of the imprecise definition of traditional feature regions. Secondly, feature seed points were screened by setting information entropy threshold and denoised by template detection method. Eventually, combined with the eigenvalues of Jacobian matrix at critical points, the extracted two-dimensional feature regions are classified, and the continuous three-dimensional flow field features are visually tracked. By analyzing the experimental results of ocean flow field data of different depth and dimension, the validity and feasibility of topological feature structure classification and tracking algorithm are proved.
{"title":"Topology-oriented 3D ocean flow field feature classification and tracking algorithm","authors":"Y. Liu, Bo Qin, Haiyan Liu","doi":"10.1145/3556677.3556683","DOIUrl":"https://doi.org/10.1145/3556677.3556683","url":null,"abstract":"The tracking analysis of ocean feature phenomena exists many problems, such as incomplete topological structure information extraction and unclear time-varying law information display, etc. In this paper, a topology-oriented 3D ocean flow field feature classification and tracking algorithm is proposed to solve the problem of flow field feature tracking in different scales. The algorithm consists of three parts: Initially, the adaptive circular sampling space manner is optimized and improved to adapt to the extraction of flow field feature regions at different scales in view of the imprecise definition of traditional feature regions. Secondly, feature seed points were screened by setting information entropy threshold and denoised by template detection method. Eventually, combined with the eigenvalues of Jacobian matrix at critical points, the extracted two-dimensional feature regions are classified, and the continuous three-dimensional flow field features are visually tracked. By analyzing the experimental results of ocean flow field data of different depth and dimension, the validity and feasibility of topological feature structure classification and tracking algorithm are proved.","PeriodicalId":350340,"journal":{"name":"Proceedings of the 2022 6th International Conference on Deep Learning Technologies","volume":"40 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-07-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127513010","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：